init

utils/semanticBoolean.ts

@@ -0,0 +1,29 @@
+import { z } from 'zod/v4'
+
+/**
+ * Boolean that also accepts the string literals "true"/"false".
+ *
+ * Tool inputs arrive as model-generated JSON. The model occasionally quotes
+ * booleans — `"replace_all":"false"` instead of `"replace_all":false` — and
+ * z.boolean() rejects that with a type error. z.coerce.boolean() is the wrong
+ * fix: it uses JS truthiness, so "false" → true.
+ *
+ * z.preprocess emits {"type":"boolean"} to the API schema, so the model is
+ * still told this is a boolean — the string tolerance is invisible client-side
+ * coercion, not an advertised input shape.
+ *
+ * .optional()/.default() go INSIDE (on the inner schema), not chained after:
+ * chaining them onto ZodPipe widens z.output<> to unknown in Zod v4.
+ *
+ *   semanticBoolean()                              → boolean
+ *   semanticBoolean(z.boolean().optional())        → boolean | undefined
+ *   semanticBoolean(z.boolean().default(false))    → boolean
+ */
+export function semanticBoolean<T extends z.ZodType>(
+  inner: T = z.boolean() as unknown as T,
+) {
+  return z.preprocess(
+    (v: unknown) => (v === 'true' ? true : v === 'false' ? false : v),
+    inner,
+  )
+}